Towards Improving the Anti-attack Capability of the RangeNet++

Qingguo Zhou, Ming Lei, Peng Zhi, Rui Zhao, Jun Shen, Binbin Yong; Proceedings of the Asian Conference on Computer Vision (ACCV) Workshops, 2022, pp. 56-67

With the possibility of deceiving deep learning models by appropri-ately modifying images verified, lots of researches on adversarial attacks and adversarial defenses have been carried out in academia. However, there is few research on adversarial attacks and adversarial defenses of point cloud semantic segmentation models, especially in the field of autonomous driving. The stabil-ity and robustness of point cloud semantic segmentation models are our primary concerns in this paper. Aiming at the point cloud segmentation model Range-Net++ in the field of autonomous driving, we propose novel approaches to im-prove the security and anti-attack capability of the RangeNet++ model. One is to calculate the local geometry that can reflect the surface shape of the point cloud based on the range image. The other is to obtain a general adversarial sample related only to the image itself and closer to the real world based on the range image, then add it into the training set for training. The experimental re-sults show that the proposed approaches can effectively improve the Range-Net++'s defense ability against adversarial attacks, and meanwhile enhance the RangeNet++ model's robustness.